I Did the Math on Kimi K3. The $15 Output Price Isn't the Whole Cost Story.

Short summary

A rigorous fact-check and cost analysis of Moonshot AI's Kimi K3 launch claims. The author corrects two false headlines — weights are not yet downloadable (planned by July 27) and 2.8T is total capacity, not active parameters per token. API pricing of $3/$15 per million tokens looks competitive, but measured verbosity of ~130M output tokens versus 63M median for comparable models can erase the per-token advantage. The author recommends task-level evaluation with output caps before making K3 a default route.

• •Kimi K3 weights are not downloadable yet — Moonshot plans release by July 27, 2026
• •2.8T is total MoE capacity, not active parameters per token; active count undisclosed
• •Verbosity measured at ~2x comparable models can negate attractive per-token pricing
• •Recommends output caps and task-level evaluation before migrating from current APIs

Generated with AI, which can make mistakes.